XAI in Brain Tumor Detection: Comparative Insights on CNNs via Gradient and Perturbation Methods

Explainable Artificial Intelligence (XAI) is essential for bridging the gap between high-performance deep learning models and their practical adoption in clinical brain tumor diagnosis. This study emphasizes the integration and comparative analysis of advanced XAI techniques—namely Grad-CAM, Grad-CAM++, saliency maps, occlusion sensitivity, and SmoothGrad—with five state-of-the-art convolutional neural network (CNN) architectures: ResNet, MobileNet, DenseNet, VGG16, and a custom CNN model. Using annotated brain MRI datasets, we evaluate not only the classification accuracy but also the interpretability of each model by generating spatially precise visual explanations that highlight tumor-relevant regions influencing predictions. We quantitatively measure the alignment of XAI heatmaps with expert tumor annotations and qualitatively assess their clinical plausibility. Our findings reveal significant variability in the quality of explanations across models and XAI methods, underscoring the critical role of XAI in validating model decisions and increasing transparency. This work contributes a rigorous framework for incorporating explainability into brain tumor detection pipelines, facilitating enhanced clinician trust and paving the way for safer AI deployment in medical imaging.